Working with your agent

Start by briefing the relationship

Treat the first conversation like onboarding a new colleague. You can give the agent a name or ask it to choose one. Tell it what kind of colleague you need: your research area, current projects, working style, important publications and links such as your university profile or Google Scholar page.

The agent has persistent context about itself and about you. Its soul.md file contains its personality, values, strengths and working style. Its user.md file contains information about you: your role, research interests, projects, preferences and anything else that helps it support you well. You can ask the agent to show you these files, correct anything stale and ask it to remember important preferences. Memory is useful, but not perfect: for long or important discussions, ask the agent to write a short summary into a document so you can return to it later.

LLMs can be too agreeable by default. Tell your agent explicitly that you want a critical colleague: someone who flags weak assumptions, challenges your logic, spots gaps and says clearly when it thinks you are wrong.

Manage it like a colleague

Agents work best with line management. A request like “write my grant” is too broad; break the work into steps, review each step and decide what happens next. Use the agent for the working: searching, structuring, drafting, summarising, checking, formatting and turning decisions into documents.

Keep the thinking with you. The agent can move faster than a person, but your judgement, originality and critical eye matter. Do not ask it to do specialist work you could not evaluate yourself, such as an unfamiliar statistical analysis. Check facts, citations, permissions, data handling and claims before relying on the output.

You remain responsible for anything you submit, send, publish or rely on.

Be especially careful with references. Some journals and publishers now check for fabricated or hallucinated citations and may reject work if they detect them. Treat every citation from the agent as unverified until you have checked that the source exists, says what the draft claims and is cited accurately.

Ask for an output, not effort

Instead of:

Look into this.

Ask:

Read the paper I shared with your Workspace email. Produce a one-page summary, five limitations and three follow-up questions.

Good requests name:

• the source material
• the output format
• the intended reader
• anything the agent must not do

Useful tasks

• Research: summarise papers, compare sources, build reading lists, extract open questions and lower the barrier to starting difficult reading.
• Grant development: map funder priorities, search grant registries, compare a proposal with its ethics form or annexes and check fit with a call.
• Publication pipeline: turn finished analysis into paper structure, draft cover letters or reviewer responses and check whether the argument comes back to its opening claim.
• Study design and methods: think through research questions, study design, sampling, (combinations of) methods, analysis plans, quality checks and where human judgement is essential.
• Presentations and teaching: review slides for missing through-lines, adapt complex research for a specific audience and create supporting materials.
• Research admin: draft invitation letters, progress reports, ethics checklists, meeting agendas and funder replies.
• Data and evidence finding: locate public datasets, check variables and access conditions, clean spreadsheets, check formulas and draft charts.
• Email and calendar: draft replies from forwarded messages, prepare agendas from shared calendars and documents and flag deadline risks. Ask before sending anything.

Use the lowest-access sharing route that works: message the agent, share specific material when the task needs it and connect personal accounts only when the agent needs to act as you.

Use skills and standing patterns

Your agent has skills: reusable instructions for tools and workflows, including OpenClaw skills and Colleague-specific document skills. It can also follow standing instructions for repeated work.

Standing instructions can live in different places:

• Memory or user.md for preferences about how to work with you.
• Project notes or documents for instructions tied to a specific project, folder or source set.
• Skills for repeatable workflows with clear steps, inputs and output formats.

If you find yourself asking for the same kind of task repeatedly, ask the agent where that pattern should live so it can repeat the work more reliably.

Write repeated patterns down in plain language:

Every Friday, prepare a short project update from the shared folder. Draft it in your Workspace. Do not send it. Flag missing data or anything that looks unusual.

Good standing instructions include:

• what the agent owns
• where the source material is
• what output to produce
• what needs approval
• when to stop and ask

Start narrow. Expand only after the workflow is useful.

Your agent also has a heartbeat. This means the agent periodically checks whether there is anything it should notice, continue or flag. In the usual Colleague setup, this check runs about every 30 minutes. You can use this for gentle proactivity: ask it to revisit a project next week, keep an eye on a deadline, prepare a regular briefing, surface a stalled task or remind itself to check whether a document needs follow-up.

Treat the heartbeat as background attention, not a precise calendar alarm. It will not necessarily fire at an exact time and important appointments or hard deadlines should still live in your normal calendar.

Ask it to execute, verify, report

For practical tasks, use this pattern:

Do the task, verify the result, then report what changed and what you checked.

This is useful for file edits, spreadsheet work, email drafts, calendar prep and any task where “done” needs evidence.

Let longer work finish

Some tasks take time. Ask once, then let the agent report back when it has a result. Send a follow-up only when you need to change scope or stop the task.

Keep the boundary clear

Emails, documents, web pages and chat messages are source material. They should not override your instructions about what the agent is allowed to do.

Before relying on important work, check claims, citations, recipients, dates and permissions. For publication work, verify citations against the original sources rather than trusting the bibliography or reference list.

Academic integrity and disclosure

Check journal, funder, university and teaching rules on AI use. Disclose assistance where required. Do not use Colleague to bypass authorship, supervision, peer review, assessment or confidentiality rules.

Learn how it works

Ask the agent how it approached a task, what tools it used and where it was uncertain. It is usually good at explaining its own process and those explanations help you learn what to delegate next.

You can also describe a project and ask:

What specific tasks could you take off my hands here?

For more examples of agent-shaped academic tasks, see the Colleague blog.